the speech to text model dropdown showed the same model twice, two rows, one file. the config stored the model as a short path without the directory, the list returned the full path with it, and nothing compared the two. the select decided the config held a model the folder did not contain, so it prepended that value as a custom row. every option is labelled by base name, so both rows read identically and it looked like two engines running. nothing was running twice.
the fix is one comparison function that matches a short path against the tail end of a full path, ignoring case and slash direction. it runs in the field builder and in the four second refresher so the duplicate cannot come back. tested against the real strings: short against full matches, full against full matches, different paths do not match, empty does not match.
while in there found the hide rules for the model list are three: the model in use is never hidden, a hidden path whose file is gone gets dropped, and picking or adding a model takes it out of the hidden set. a freshly added model was invisible because earlier hidden clicks had crossed it out. the backend was answering correctly the whole time, the list was refusing to draw it.